aquif-3.5-Max-1205

The latest iteration of the aquif-3.5-Max series, representing frontier-level reasoning and problem-solving capabilities. aquif-3.5-Max-1205 builds upon the proven architecture of aquif-3.5-Max with targeted improvements in reasoning depth, mathematical performance, and code generation quality.

Model Overview

Attribute	Value
Total Parameters	42.4B
Active Parameters	3.3B
Architecture	Mixture of Experts
Context Window	1M tokens
Reasoning	✅ Thinking-Only
Precision	BF16/FP16
License	Apache 2.0

Key Improvements

aquif-3.5-Max-1205 introduces enhanced reasoning through advanced RL and SFT techniques, delivering measurable improvements across:

Advanced Reasoning: Deeper problem decomposition and logical inference
Mathematical Excellence: Superior performance on competition-level mathematics
Code Mastery: Enhanced code generation, analysis, and optimization
Long-Context Handling: Stable reasoning across full 1M token window

Evaluation

Coding evals

Artificial Analysis

Benchmark	aquif-3.5-Max	aquif-3.5-Max-1205	Improvement (%)
MMLU-Pro	85.4	85.2	-0.23%
GPQA Diamond	83.2	82.7	-0.60%
AIME 2025	94.6	95.0	+0.42%
LiveCodeBench	81.6	84.1	+3.06%
Humanity's Last Exam	15.6	17.3	+10.90%
TAU2-Telecom	51.3	62.7	+22.23%
IFBench	65.4	66.3	+1.38%
TerminalBench-Hard	23.9	28.5	+19.25%
AA-LCR	61.2	62.4	+1.96%
SciCode	40.9	42.3	+3.42%
AA-Index	60 (60.31)	63 (62.65)	+5.00%

Technical Specifications

Mixture of Experts Architecture: 42.4B total / 3.3B active parameters
Reasoning Modes: Toggle between thinking and non-thinking modes
Attention Mechanisms: Optimized KV caching and efficient sparse routing
Multilingual: 10 languages (EN, DE, IT, PT, FR, HI, ES, TH, ZH, JA)
Precision Support: BF16 and FP16

Usage Recommendations

Frontier-level problem-solving without compromise
Advanced research and scientific computing
Competition mathematics and algorithmic challenges
Comprehensive code analysis and generation
Complex multilingual tasks requiring maximum reasoning capability
Extended document analysis and synthesis

Thinking-Only Architecture

aquif-3.5-Max-1205 operates exclusively in thinking mode. Unlike its predecessor, this model cannot disable reasoning—the RL pipeline ensures that extended reasoning is always active. This design choice prioritizes maximum depth and capability for frontier-level problem-solving, with thinking patterns fully integrated throughout the model's inference.